We are now accepting applications for our prestigious postdoc program. 3 year appointments. Flexible mentorship. No teaching requirement. $100,000/yr + $10k/yr in travel and research funding. Excellent benefits. We’re looking for the best. Join us at FDS.

If you’re looking to help people impacted by the floods in Libya, check out these organizations providing relief:

This will be presented today at #NeurIPS (11am-1pm) Poster #1036. Come through if interested in boosting robustness to adversarial examples!

Joint work with Avrim Blum, Greg Shakhnarovich, and @hongyangzh. I hope you enjoy the read!

Our results reveal that the two problems of barely robust learning and strongly robust learning are actually equivalent. There is also an interesting landscape for boosting robustness that emerges with connections to the classic and pioneering works on boosting the accuracy.

Our formalized notion of "barely" robust learning requires robustness with respect to a "larger" perturbation set, which we show is *necessary* and that weaker relaxations such as robustness with respect to the actual perturbation set that we care about is *not* sufficient.

Motivated by this, we study the theoretical question of boosting "barely" robust learning algorithms, and we provably show that it is possible to boost their robustness with a novel boosting algorithm.

Adversarially robust learning has been quite challenging in practice. Our current algorithms are able to learn predictors with low natural error but robust only on a small fraction of the data distribution (sometimes even less than 50%).

Can we boost barely robust learning algorithms to learn predictors with high robust accuracy? I am very excited to share new work putting forward a theory for boosting adversarial robustness: https://t.co/PLioXF04yX. (1/6)

Thank you @KPCBFellows for the wonderful gift! @justinsayarath your tip was on the spot, I love gelato :D http://t.co/mMSOuxnxl6